The big data of bad driving, and how insurers plan to track every turn

Discriminating against customers on the basis of driving habits that they have control over, like how fast they drive, or how hard they swerve, is completely fine.

If driver A drives safely at the speed limit, and driver B constantly pulls stunts at speeds well over the speed limit, B is much more likely to get into an accident than A. Why should A have his premiums raised just to subsidize B's voluntary driving habits?

By making B pay for the costs of his unsafe driving habits, and rewarding A for his safe driving habits, this feels like a great marketplace improvement.

It will be sold/marketed as a discount if you use it, of course, but it's the same.

It remains to be seen whether it will have any actual effect on the way people drive.

Alternatively, self-driving cars take over before we're are effectively forced to use tracking

For example, is driving more aggressively more dangerous than texting (or simply not being attentive to the road)? Because only the former is tracked.

Car (really, driving) insurance is heavily regulated by the government, each state has agency over what information insurers are allowed to include in their risk models, as well as actually auditing the risk models themselves. Insurance nomenclature will refer to these inputs as risk factors. You can get an overview of the rules in California from the Insurance Commissioner's office here [1], and you should be able to find similar overview for your state by googling "auto insurance risk factors <state>"

Usage-based insurance (UBI) is the most recent innovation to be approved for widespread use. It's actually been around for a few years (at least 2012), this article's relevance is mostly about the incorporation of UBI data-gathering sensors directly into new vehicles. The key risk factors UBI focuses on are miles driven, and extreme acceleration/deceleration events. These factors were chosen because they can be easily measured using existing data ports on older vehicles [2], amongst other reasons. The San Francisco-based start-up Metromile [3] is one of the industry leaders.

Ultimately, UBI is limited in that it focuses exclusively on factors specific to the vehicle. As many commenters have identified, "risky" driving behavior is a both a function of how you drive, and how the people around you are driving. The latter effect is much, much harder to capture, but the US Department of Transportation is already engaged in R&D to identify potential risk factors that could be used within the existing regulatory framework.

With the rise in "semi-professional" drivers via on demand ride-sharing services, changes in ownership models via car sharing services, and the rise of semi-autonomous and eventually autonomous vehicles, the mechanics of automotive insurance are going to change a lot in the next few years.

Happy to answer questions as best I can if people are curious.

[1] http://www.insurance.ca.gov/0400-news/0200-studies-reports/0...

[2] https://en.wikipedia.org/wiki/On-board_diagnostics

[3] https://www.metromile.com/

If you enable a random insurance company to have all data about where exactly you have been every moment of the day, you are giving up a lot of your privacy. And I bet the terms of service will be the standard "Oh, by the way, we are going to share this with all our partners" kind.

Now the exception to above is perhaps a world where none of these autodrives are privately owned. One might think that in such situations a passenger's movements would be anonymous. They won't be. Even cabs/ubers today keep track of who goes where. Whenever a means of transport is not privately owned, payment will be an issue. And payment leads to individualized tracking.

The further exception to above would be autodrives owned by google and paid for with untraceable cryptocurrency (or cash) ... that's one too many layers of fantasy to debate.

With pilots we have minimum flight time per year as well as medical requirements. I suggest similar but for cars because people physically unable to drive or people just being morons driving 10 km/h under the speed limit are a risk to everyone else, not just them.

The problem is that the trend to data-driven insurances and other bullshit like this won't help for these cases because computers and especially deep learning models are inherently fallible. A skilled mathematician may look at a risk calculation and check if the premium is adequate (and in case of old people needing their cars to not lose their entire freedom, also "close his eyes")... no way to do so with a neural network. Not to mention that all that data IS going to be abused. If you want to check for public acceptance of government snooping, look at the fan reactions for NCIS episodes. No one ever objects to the portrait that surveillance and hacking is neccessary for law enforcement...

How does the insurance company know you are hard braking because you are an an inattentive jackass that is texting while driving that suddenly has to brake to avoid a collision vs a safe driver that had to suddenly account for someone in front of them making an improper lane change?

The short answer is they can't but you can bet that they'll use that data to justify raising your rates.

https://www.kaggle.com/c/axa-driver-telematics-analysis

Also, of course, the insurance companies have different incentives than I do as a driver, or, indeed, society as a whole. The the stereotypical clueless/passive/elderly/(dare I say Prius) driver who accelerates incredibly slowly, brakes incredibly slowly, vaguely lists from lane to lane (without signaling) is a terrible driver, and is arguably dangerous to everyone around them. However, they might move so slowly that other drivers are able to anticipate and avoid them, so the accident claims data might not show them as high risk. (although they might tend to make lots of parking lot claims as they 'feel' their way into a parking spot). As a society we don't want to encourage clueless driving; on the other hand (if they can avoid parking lot scrapes), the insurance company might well prefer this driver.

The insurance claims-based model absolutely incentivizes attentiveness, as you can avoid both your own accidents, as those that others attempt to cause. It's certainly not perfect (particularly when paired with things like credit checks), but I'd hate to see drone-like driving preferred over attentive driving.

Another thing to note is that folks with an active interest in driving (engaged, sometimes drive fast, have advanced training, etc) will absolutely avoid these technologies because they're all convinced they'll be punished in the model -- so even if the insurance companies are actively collecting data in an attempt to refine their models, it will have a selection bias.

The main thing these data will be lacking is context. Did I brake hard because I was texting and almost hit the car in front of me? Or did I brake hard because I goddamned felt like it? Were there cars behind me at the time, or was I driving in such a manner that there was no risk of collision? Can it separate inputs that are likely to cause two vehicles to collide, versus inputs that might cause a single-vehicle collision (in which case context doesn't matter)?

Hell, just last week I unexpectedly threw my brand-new car into a full ABS panic stop to avoid some deer. Then again, I guess the insurance company doesn't care why. It's better than a collision, but they'd prefer that I never drive anywhere I might ever get in an accident. Actually, they'd just prefer I don't drive anywhere at all.

What I was aware of during the tracking process was how little the app was aware of - if was sensing acceleration and GPS position, but nothing else, and given there were plenty of times that it wasn't aware of the on road situation (such as someone pulling out in front of me, a road blockage, traffic, etc), I was thinking that it really can't tell everything it needs to - or indeed if I put my indicators on when I change lane, etc. Even if you had a system that connected to this CAN bus to sense such things, the picture would be incomplete.

However, I'm sure that it will be useful to a degree, but there will be plenty of outliers in the data where there are safe drivers who are marked as unsafe and vice versa.

If you can convince people to regulate their own behavior the way you want, all the better. This total surveillance society we're building means more people are looking over their shoulders and not wanting to "get caught" doing something someone else doesn't like. It's pure profits for the plutocrats.

Based on the way I see people drive in Hampton Roads, VA I don't think any one would ever use this. There are maniacs at every turn. If you are willing to risk going the speed limit on I-64/262/464 you're likely to get run over.

But I'm sure law enforcement will appreciate another massive database.

Business is allowed to use any input variable in the whole scheme to jack up prices. Business is not mandated to lower them for any reason.

Well, lets give them more input on how to jack up more prices.

This conversation has been had in Ontario, Canada for so long where the insurance rates are the highest in all of Canada. It is not unheard of to pay 5000 to 7000 in Brampton, Ontario. If you move over just to the west, your rates will drop by 2000.

Laws need to change, as usual.

The problem would be finding the right size and composition of people that you can somewhat trust to be reliable drivers and that use their car around the same amount as you so that the risk profiles are all roughly the same.

so I thought the actual role of insurance is to cover the case where the difference from the arithmetic average cost (i.e. average of all cases, consisting overwhelmingly of neutral cases, but with some rare negative possibilities) and the negative event is so large that the rare negative event would be life-changing but the average case is trivial.

for example, it might make total sense for a tiny startup to take out a policy against a total, multi-day AWS outage. This is because that event is basically impossible to occur, but if it did it might bankrupt them immediately. (You can use your imagination as to why this might be.)

So the basic role of insurance is to reduce the variation - get things closer to the arithmetic average case.

But does this work if people don't know the average case?

It is easy to find nationwide accident rates, one could make an educated guess about it. But if insurance companies are tracking your every turn and coming up with a unique offer - what is to say it is remotely fair?

How can anyone tell if their insurance comapny, "You, personally, have driving habits that give you a 10% chance of having an accident this year, and it's remarkable that you haven't had any in the past 8 years you've been driving. You've simply been lucky."

If they're a big-data insurance company, who are you to argue? After all, your friends have been in accidents, you know it's risky.

What I'm getting at is there is huge information-asymmetry if the insurance company has access to a lot more information than the person buying the insurance does!

What's to keep them from selling you tsunami-insurance at Japanese or Californian rates, even though you live in Minnesota?

In the status quo, it's because you KNOW that you live in Minnesota. But if your every choice is being tracked, doesn't that assumption go out the window?

Of course, some kind of a legal requirement to share their actuarial calculations with you would never work, because there is not a company in the world that wouldn't figure out how to lie - you can't get companies to share their profit margins or to keep to a certain profit margin, it's just literally impossible. (It would be like getting your SO not to think of anyone else during sex, no matter how fleetingly. That is just not a possible goal, and I use this analogy because companies' internal calculations are similar to what is going on in someone's head.)

So if private companies can't be held to a profit margin, is the answer state-issued insurance? This would be beneficial in many ways: for example, if someone's behavior is so risky that their average case is pretty bad, then perhaps the state can still choose to subsidize them from less risky customers. A private company might choose to fire them as a customer (or offer them an unaffordable rate.)

The difference is, when prices are close to unaffordable, can insurance companies mislead customers into thinking that it is because they, personally, really are that risky?

How do we stop this? I think the answer is to simply not allow this level of personalization.